- [pdf] [supp] [code]
Emphasizing Closeness and Diversity Simultaneously for Deep Face Representation
Recent years have witnessed remarkable progress in deep face recognition due to the advancement of softmax-based methods. In this work, we first provide the analysis to reveal the working mechanism of softmax-based methods from the geometry view. Margin-based softmax methods enhance the feature discrimination by the extra margin. Mining-based softmax methods pay more attention to hard samples and try to enlarge their diversity during training. Both closeness and diversity are essential for discriminative features learning; however we observe that most previous works dealing with hard samples fail to balance the relationship between closeness and diversity. Therefore, we propose a novel approach to tackle the above issue. Specifically, we design a two-branch cooperative network: the Elementary Representation Branch (ERB) and the Refined Representation Branch (RRB). ERB employs the margin-based softmax to guide the network to learn elementary features and measure the difficulty for training samples. RRB employs the proposed sampling strategy in conjunction with two loss terms to enhance closeness and diversity simultaneously. Extensive experimental results on popular benchmarks demonstrate the superiority of our proposed method over state-of-the-art methods.